Ensemble machine learning models for monitoring riparian vegetation dynamics using historical aerial orthophotos

Historical aerial photographs are well-known as a reliable source of information on historical land cover and land use. However, extracting this information can be challenging due to the limited spectral characteristics in black-and-white images. In this study, we evaluate a textural-based approach using Machine Learning (ML) models to detect the spatial pattern of the braided-wandering multichannel system from historical aerial images with an emphasis on riparian vegetation.

Five aerial datasets (1949–1992) were used to extract textural information through Gray level Co-occurrence Matrix (GLCM) and geomorphological operations on High-resolution, preprocessed, and normalized orthophotos. We used Random Forest (RF), Light gradient boosting machines (LightGBM), and Extreme Gradient Boosting (XGBoost) ML methods through two classification schemes to classify images into five main classes. GridSearchCV hyperparameter optimization tool were utilized to optimize models and Sequential Feature Selection (SFS) algorithm to reduce the dimensionality of the data cube. The results indicated the efficacy of Morphological operations (Gradient, Eroded, and Dilated) and GLCM features (contrast, entropy) in the final classified map. The RF model demonstrated greater stability and higher median accuracy across datasets. While there was no significant difference between LightGBM and XGBoost in terms of accuracy metrics, XGBoost's performance was notably more variable but significantly faster. In our study, the shadow effects, distortion, and radiometric differences within the orthophotos remain challenging. Despite limitations, the proposed approach addresses key challenges in extracting information from historical orthophotos and can be extended to broader ecological and environmental applications.